fix: ggml_backend_cuda_get/set_tensor_async — fall back to buffer's own method when tensor is non-CUDA

When the scheduler calls get_tensor_async(cuda_backend, tensor) but the tensor's
buffer is CPU/other, delegate to the buffer's own get_tensor/set_tensor instead
of asserting. This fixes the 'unsupported buffer type' crash when the scheduler
tries to read tensors through the wrong backend.

Author: Alex7MV

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -3261,7 +3261,11 @@ static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tens
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    GGML_ASSERT(ggml_backend_buft_is_cuda(buf->buft) && "unsupported buffer type");
+    if (!ggml_backend_buft_is_cuda(buf->buft)) {
+        buf->iface.set_tensor(buf, tensor, data, offset, size);
+        return;
+    }
+
     ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buf->buft->context;
     GGML_ASSERT(buft_ctx->device == cuda_ctx->device && "buffer device mismatch");
 
@@ -3272,12 +3276,13 @@ static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggm
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    GGML_ASSERT(ggml_backend_buft_is_cuda(buf->buft) && "unsupported buffer type");
-    ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buf->buft->context;
-    if (buft_ctx->device != cuda_ctx->device) {
-        fprintf(stderr, "CUDA get_tensor_async: buf dev=%d ctx dev=%d tensor=%s size=%zu\n",
-            buft_ctx->device, cuda_ctx->device, tensor->name, size);
+    if (!ggml_backend_buft_is_cuda(buf->buft)) {
+        // Tensor not on CUDA — use buffer's own get_tensor
+        buf->iface.get_tensor(buf, tensor, data, offset, size);
+        return;
     }
+
+    ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buf->buft->context;
     GGML_ASSERT(buft_ctx->device == cuda_ctx->device && "buffer device mismatch");
 
     cudaStream_t rb_stream = cuda_ctx->readback_stream_get();

src/llama-context.cpp

@@ -2094,8 +2094,6 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
                             ggml_backend_sched_reset(sched.get());
                             for (int i = 0; i < ggml_graph_n_nodes(phase2_gf); i++)
                                 ggml_backend_sched_set_tensor_backend(sched.get(), ggml_graph_node(phase2_gf, i), gpu);
-                            for (int i = 0; i < ggml_graph_n_leafs(phase2_gf); i++)
-                                ggml_backend_sched_set_tensor_backend(sched.get(), ggml_graph_leaf(phase2_gf, i), gpu);
                             force_idxs_to_cpu();
                             if (!ggml_backend_sched_alloc_graph(sched.get(), phase2_gf)) {
                                 ret = GGML_STATUS_ALLOC_FAILED; return nullptr;